Deep learning techniques for the study of student's academic performance during distance education caused by COVID-19

In this work, we study the influence of various educational factors on the academic performance of school students during the transition to distance learning format due to the COVID-19. The coronavirus pandemic affected not only health system, but also all aspects of human lives, including educational system as well. The study is based on a large dataset consisting of school students marks for 2020 academic year taken from “Electronic education in Tatarstan Republic” system's database. The data analysis and interpretations were performed by using modern deep learning methods (feed-forward neural networks) and feature importance detection techniques by using Captum framework. This method allows us not only to obtain features with the most significant impact on the prediction, but also introduces numerical characteristics for its quantitative evaluation and comparison. The studied educational factors include teachers' characteristics, mean marks for pre-pandemic period, number of grades, group of school subject and school characteristics. As a result, we discovered that the most significant feature is the mean mark for pre-COVID19 period, thus, general trend of school marks remained pretty similar to that period, but not identical. Also, we established importance on pupil's changes of marks a certain schools and types of schools, teachers with certain age and gender that can be explained by different period of adaptation to the rapidly changing situation.